The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and parses README files from arbitrary Hugging Face repositories using the extract-readme command. If a repository contains a malicious README, it could attempt to influence the agent's behavior during table extraction or subsequent evaluation steps.
Ingestion points: scripts/evaluation_manager.py (via extract-readme and inspect-tables).
Boundary markers: The skill uses token normalization and markdown parsing to identify tables, but lacks explicit "ignore embedded instructions" warnings when processing README text.
Capability inventory: File writing (metadata updates), PR creation, and job execution via hf jobs.
Sanitization: The skill performs normalization of model names (tokenization, lowercase, stripping markdown) but does not sanitize the full content of the table against natural language instructions.
[COMMAND_EXECUTION]: The skill relies on executing local Python scripts via uv run to perform its core functions. It also provides commands to submit these scripts to Hugging Face infrastructure via hf jobs uv run.
[DYNAMIC_EXECUTION]: The documentation explicitly mentions the --trust-remote-code flag for models with custom architectures (e.g., Phi-2, Qwen). This flag enables the execution of arbitrary Python code defined within a model's repository during the loading process. Additionally, uv run utilizes PEP 723 script headers to dynamically resolve and install dependencies at runtime.

hugging-face-evaluation